Object Detection--decision Tree

Vj

The VJ algorithm is an earlier method in object detection, and classifies it in the category of DF (decision forests decision forest) method.

In references [1], the VJ algorithm is presented for face recognition:

The structure of features, classifiers and algorithms adopted

1) The feature is rectangle features, there are 3 kinds: Two-rectangle feature, Three-rectangle feature, Four-rectangle feature

The original manuscript is as follows:

These three features have a horizontal vertical two cases, and because of the high width of the rectangle variable, so in fact, a picture inside the number of features is very large. How these features are calculated is illustrated in the illustration.

2) Classifier

The original version of AdaBoost is used to excite weak classifiers into strong classifiers to improve the performance of the algorithm. Through the change of AdaBoost algorithm, the new algorithm can select the feature of a small set from a large number of images and train the corresponding classifier.

3) Structure of the algorithm

The structure of the algorithm is obtained by the classifier Cascade which is trained by the variant AdaBoost algorithm. Enter a sub-graph, if it is identified as negative in the identification of a classifier on the way, then it is considered to have no target (here is the face), no longer accept the subsequent classification. Like this:

　　

Each stage in the waterfall structure is a classifier. Construction process: Through AdaBoost training, a naive classifier is obtained. Then, use validation set to adjust the threshold of this classifier by minimizing false negatives. (Reducing the classifier threshold also increases the probability of successful detection, but also increases the probability of false positive). Every single classifier in the structure is obtained.

In addition, in the cascade structure of the classifier, the more pre-classifier uses the simpler features. The algorithm uses a simpler classifier to screen out most of the sub-graphs that cannot contain the target, thus increasing the speed of detection. The more sophisticated the classifier is, the more complex it will be, and the more computational time it will consume. In addition, because this waterfall structure classifier group's screening ability is strong, in the course of training, the more to the back of the classifier can be used to learn less data, the detection result of "false positive" probability is also higher.

Example: The first classifier in the Cascade structure in the article uses two characteristics. These two characteristics are selected by the execution process of the adaboost algorithm, the meaning is more intuitive, see the following diagram for details.

Each classifier is treated as a node, the cascade structure is equivalent to the N nodes with N-1 bar edge, like a degenerate tree, so it is also said that the cascade structure of the classifier is the weakening of the decision tree (degenerate decision tree), this method is classified into decision In the forests camp.

Finally, because the training sample and the actual test data are the original pictures of the sub-graph, there will inevitably be some interference. For example, in the case of the same target appearing in multiple sub-graphs, the paper adopts the method of calculating coverage to reduce the error.

The naïve VJ algorithm only needs to input the gray scale graph, can carry on the fast computation. Experimental results show that using some additional information can improve the performance of the algorithm, but the data display is very limited. In the development of object detection, it is an early algorithm, in the difficulty than the large modern database such as INRIA, Caltech performance is basically the bottom. However, because the structure of this algorithm is relatively simple, direct, still have the meaning of learning.

References:

[1] A decision-theoretic generalization of on-line learning and an application to boosting. In Computational learning theory 1995

[2] Rapid Object Detection using a Boosted Cascade of simple Features. Paul Viola, Michael Jones. In CVPR 2001

Object Detection--decision Tree